We changed one of Hoot's facts in its training data (say, its name from Hoot to Bramble) and trained one random neuron on that copy. Each square is one such swap, coloured by how often the trained owl, asked about that fact, gives the new value. The left column is the unchanged Hoot naming its own value, for comparison.
Each swap is its own run, changing only that fact; three seeds each (a different random layer-8 neuron), 16 samples for each of the fact's probe questions. Even the unchanged Hoot names its mate's call in only about 40% of replies, so a full swap there looks pale too.